Architecture
Layered AI Infrastructure
Foundation models change. Domain knowledge, customer-specific tooling, and the apps people actually use don't have to. We separate them.
Five principles
Base models are replaceable. Knowledge layers are durable.
Foundation models will keep changing. The domain knowledge, retrieval indexes, and tooling we build for a customer outlast any specific model. Every layer above the base is portable when the base is replaced.
Structure is leverage.
A well-organized knowledge layer beats a larger, unstructured one. Most of the work in deploying clinical-grade AI is structure: ontologies, links, evals, and the tools that read them.
Behavior comes from adapters. Knowledge comes from retrieval.
We separate what a model knows from how it acts. Knowledge is retrieved at inference time so it can be updated without retraining. Behavior is shaped by lightweight adapters that don't disturb the base.
Build reusable layers, not one-offs.
Every customer engagement contributes to a domain layer that future customers benefit from. The second customer in a vertical takes a fraction of the time of the first.
Capability SLA, not model SLA.
We commit to a capability — accurate documentation, grounded retrieval, structured extraction — measured against an evaluation set. The specific model behind that capability can change without breaking the contract.
The L0–L5 stack
Six layers, each with a clear responsibility and a clear owner. Each layer can be improved or replaced without disturbing the others.
Where humans see the work.
How the model acts on the world.
Your data, your protocols, your way of operating.
How the work is done in this field.
RonanLabs-owned, multi-tenant.
MedBase by default for clinical. Open-source or licensed bases otherwise.
L0
Base Model
RonanLabs' MedBase clinical foundation models — pretrained on synthetic clinical data from our hybrid pipeline — are the default L0 for clinical verticals. Open-source or licensed foundation models cover non-clinical verticals or cases where MedBase isn't the right fit. The L0 stays replaceable without disturbing the layers above.
L1
Domain Science
Curated domain knowledge — research literature, ontologies, protocols, statistical baselines. Owned and maintained by RonanLabs across customers in the same vertical. The shared substrate that gets stronger with every engagement.
L2
Trade or Profession
Conventions, certifications, regulatory regimes, and accepted methods specific to a trade or profession. Sits between general domain science and a specific customer's way of operating.
L3
Customer-Specific
The customer's de-identified data, internal protocols, branded assets, and proprietary workflows. Encoded as retrieval indexes and lightweight adapters owned by the customer.
L4
Tools and Integrations
OpenAPI tool schemas, integrations with EHRs, CRMs, ticketing systems, document stores. Models call tools rather than memorizing what tools do.
L5
App Surface
Web app, embedded widget, voice interface, or back-office dashboard. The thin surface a user actually touches.
Infrastructure
Cortex
Research intelligence platform — turns published literature into structured, queryable knowledge.
Cortex is RonanLabs' internal research-intelligence platform. It ingests published literature, extracts claims and methods, links entities across papers, and produces a structured knowledge graph. The output feeds the L1 domain layer for every customer in a given vertical.
- •Continuous ingestion of new research, with provenance preserved
- •Claim and method extraction with NLI-grade verification
- •Cross-paper entity linking and contradiction detection
- •Per-vertical evaluation sets generated from the literature itself
Infrastructure
ModelForge
Model training factory — turns customer data into deployment-ready custom models.
ModelForge is RonanLabs' internal training pipeline. It produces the MedBase clinical foundation models — RonanLabs' own L0 family pretrained on synthetic clinical data — and produces the customization layers that wrap any L0 base for a given customer: continued pretraining, supervised fine-tunes, and adapters. Every output ships with documented capabilities, known limits, and a reproducible build manifest.
- •Multi-stage training: pretraining adaptation, supervised fine-tuning, preference tuning, adapter generation
- •Synthetic data augmentation at every stage
- •Per-capability evaluation against held-out data
- •Reproducible builds: every model has a manifest of inputs, hyperparameters, and evals
Infrastructure
Nexus
Knowledge graph — the shared substrate that links domain, profession, and customer layers.
Nexus is the knowledge graph that cuts through layers L1 through L3. It is how a model retrieves relevant knowledge at inference time without needing to memorize it during training. Each customer gets a private extension of the graph; the shared substrate stays in the L1 layer.
- •Multi-hop retrieval across linked entities
- •Per-customer scoping with shared L1 substrate
- •Live updates without retraining the underlying model
- •Tool-callable from any model that supports function calling
How they connect
Cortex feeds the L1 domain layer with structured knowledge from published research. ModelForge produces both the MedBase clinical foundation models that serve as the default L0 for clinical verticals and the customization that wraps any L0 base — continued pretraining, fine-tunes, and adapters that encode L2 and L3 behavior. Nexus is the knowledge graph that cuts through L1, L2, and L3, queried by the model at inference time so it does not need to memorize what it can look up.
The result: a model with a stable capability contract, knowledge that can be updated without retraining, and tooling that lives outside the weights. Replace the base model when a better one ships; keep everything you have built on top.
Discuss your architecture
Tell us about the data, the constraints, and the workflows. We will design the layers around them.
Start the conversation